Jun-ichi Takahashi

Jacobian approach to fast acoustic model adaptation

This paper describes a Jacobian approach to fast adaptation of acoustic models to noisy environments. Acoustic models under a noise assumption are compensated by Jacobian matrices with the difference between assumed and observed noise cepstra. Detailed mathematical formulation and algorithm derivation are presented. Experiments showed that when a small amount of training data is given, this approach outperforms the existing approaches (such as PMC and NOVO) for composing a model from speech and noise models. It drastically reduces computational cost by replacing the complicated computation of model composition by simple matrix arithmetic and enables real-time environmental noise adaptation. Combination with spectrum subtraction is also discussed.

Vector-field-smoothed Bayesian learning for incremental speaker adaptation

The paper presents a fast and incremental speaker adaptation method called MAP/VFS, which combines maximum a posteriori (MAP) estimation, or in other words Bayesian learning, with vector field smoothing (VFS). The point is that MAP is an intra-class training scheme while VFS is an inter-class smoothing technique. This is a basic technique for on-line adaptation which will be important in constructing a practical speech recognition system. Speaker adaptation speed of the incremental MAP is experimentally shown to be significantly accelerated by the use of VFS in word-by-word adaptation. The recognition performance of MAP is consistently improved and stabilized by VFS. The word error reduction rate achieved in incrementally adapting a few words of sample data is about 22%.

Vector-field-smoothed Bayesian learning for fast and incremental speaker/telephone-channel adaptation

Abstract This paper describes an on-line adaptation method that combines maximuma posteriori(MAP) estimation for intra-class training (the training scheme incorporates new training samples with prior information) with vector field smoothing (VFS) for inter-class smoothing. Results of experiments comparing recognition performance of MAP/VFS with MAP adaptation for speaker adaptation and simultaneous adaptation of speaker and telephone channel show that fast and incremental adaptation can be achieved even with a relatively small number of training samples (under 10 words) due to VFSs ability to consistently enhance MAP adaptation. High word error reduction rates, which in the experiments were 22% for speaker adaptation in a large-vocabulary isolated-word recognition task (vocabulary size=2575) and 48% for simultaneous adaptation of speaker and telephone channel in a 100-isolated-word recognition task, can be achieved through word-by-word incremental adaptation using 10-word data.

ASR and TTS telecommunications applications in Japan

Abstract This paper first describes recent trends of ASR and TTS telecommunications applications in Japan. ASR applications focus on public services such as operator automation, operator assistance, voice-activated information retrieval, and voice dialing. Major TTS applications include information service by voice and e-mail reading. The usage of ASR and TTS functions is expected to dramatically increase in the near future with the penetration of handy and mobile telephone terminals; hot topics are text broadcasting and digital communication. Secondly this paper describes NTTs experimental interactive system featuring (1) highly accurate speaker independent and large vocabulary speech recognition based on context-dependent accurate acoustic phoneme HMM models trained with speech data from more than 10,000 speakers collected over telephone network, (2) high quality text-to-speech synthesis that generates speech by concatenating triphone-context-dependent waveform segments, (3) software-based configuration that requires no special hardware except a PC equipped with a sound board and a voice modem, and (4) easy and rapid prototyping which enables the developer to build a system by writing some types of service scenarios.

Interactive voice technology development for telecommunications applications

Abstract This paper describes the essential speech processing techniques for interactive voice applications in the telecommunications field. These techniques include speech recognition and speech synthesis, both of which aim to make interactive speech communications between man and machine more natural. Keyword spotting, background noise effects reduction, and speakers and/or telephone adaptation techniques are considered essential in speech recognition in order to allow a more natural voice input as well as an adequate robustness against environmental variabilities. In the area of text-to-speech synthesis, we propose a rule-based synthesis method applicable to the Japanese language, aiming to produce high quality speech. The commercial system ANSER of a former project is also described as an example of an interactive speech processing system. Finally, a recently developed speech recognition server which includes a vocabulary-flexible recognition function is described. It is meant to illustrate the concept of the techniques it employs which allow its range of applications to be easily extended and also allow it to adapt itselt to the changes which are rapidly occurring in the telecommunications field.

Iterative unsupervised speaker adaptation for batch dictation

Describes an automatic batch-style dictation paradigm in which the entire dictated speech is fully utilized for speaker adaptation and is recognized using the speaker adaptation results. The key point is that the same speech data is used both for recognition as the target and for speaker adaptation. Two steps, speech recognition and speaker adaptation which uses the recognition results as means of supervision, are iterated to maximize the advantage of closed-data speaker adaptation. Recognition errors are reduced by 37% in a practical application of batch-style speech-to-text conversion of recorded dictation of Japanese medical diagnoses compared to speaker-independent recognition. To select only reliable recognition results, a supervision improvement procedure is used, by which erroneous recognition results can be eliminated from the supervision. In this procedure, 59-74% of the data are extracted from the tentative recognition results, and their reliability is 89-93%. This procedure also reduces recognition errors by 45%.

Minimum classification error training for a small amount of data enhanced by vector-field-smoothed Bayesian learning

This paper describes an efficient method of attaining the highest level of recognition performance yet achieved in minimum classification error (MCE) training for a small amount of data. This method combines MCE and vector-field-smoothed Bayesian learning called MAP/VFS. In the proposed method, the training capability of MCE in robust acoustic modeling can be significantly enhanced with MAP/VFS. In the method, MCE training is performed using an initial model trained through MAP/VFS. The same data are used in both training. For speaker adaptation using 50-word training data, the error reduction rate drastically rises to 47% compared with 16.5% when using only MCE. This high rate, in which 39% is due to MAP, an additional 4% is due to VFS, and a further improvement of 4% is due to MCE, can be attained by enhancing MCE training capability with MAP/VFS.

Fast telephone channel adaptation based on vector field smoothing technique

The paper presents a fast telephone channel adaptation method of MAP/VFS with a sequential training function. The concept is based on using maximum a posteriori (MAP) estimation as an intra-class training scheme in combination with vector field smoothing (VFS) technique as an inter-class training scheme. Experimental results of simultaneous adaptation to a telephone channel and a speaker show the proposed method is significantly superior to sequential MAP adaptation. The error reduction rate achieved in sequentially adapting a few words of sample data is about 41% using the proposed method, while that of the sequential MAP adaptation hardly improved even with ten-word adaptation data. MAP/VFS, with its fast and sequential adaptation function, is expected to be very useful in developing telephone applications such as information services proceeded by iterative tree-structured item selection.<<ETX>>

Low out-gassing organic spin-on hardmask

Beyond 45nm node processes, ArF hyper-NA immersion lithography systems are an inevitable choice for obtaining smaller patterns. A hyper-NA, dual BARC system is proposed to achieve low reflectivity. However, the ability for the resist to ask as a mask is severely challenged because of the increased film thickness associated with a dual BARC system. In order to obtain enough etch selectivity to the substrate, multi-layer resist processes can be applied. General multi-layer resist processes uses silicon containing an inorganic spin-on hard mask and an organic spin-on hard mask with a high carbon content. One of the problems of organic spin-on hard masks is high out-gassing, which can cause defect issues in mass production. We have developed a new organic hard mask with low out-gassing, good reflectivity control (< 0.2%) and good etch durability. Gap-filling performance also can be controlled by changing its fluidity and wettability on the substrate.

High Si content BARC for applications in dual BARC systems such as tri-layer patterning

This work discusses the requirements and performance of Honeywells middle layer material, UVAS, for trilayer patterning. UVAS is a high Si content polymer synthesized directly from Si containing starting monomer components. The monomers are selected to produce a film that meets the requirements as a middle layer for trilayer patterning and gives us a level of flexibility to adjust the properties of the film to meet the customers specific photoresist and patterning requirements. Results of simulations of the substrate reflectance versus numerical aperture, UVAS thickness, and under layer film are presented. Immersion lithographic patterning of ArF photoresist line space and contact hole features will be presented. A sequence of SEM images detailing the plasma etch transfer of line space photoresist features through the middle and under layer films comprising the TLP film stack will presented. Excellent etch selectivity between the UVAS and the organic under layer film exists as no edge erosion or faceting is observed as a result of the etch process. The results of simulations of Rsub versus NA, and the thickness of each film comprising a two layer antireflective film stack will also be discussed.